← Back to feed

Thomson Reuters launched Thomson 1.0 on August 24 — the company’s first proprietary LLM, built on Qwen 3.5 and trained exclusively on its own content. The model is now in production powering the tabular analysis feature inside CoCounsel Legal.

The build

Thomson was not trained from scratch. The company started with Qwen 3.5, an open-weight model from Alibaba’s Tongyi Lab, then applied a multi-stage fine-tuning process: first aligning the model’s values and behavior to Thomson Reuters standards, then training exclusively on Westlaw, Practical Law, Checkpoint, and Reuters news archives. A final reinforcement learning phase trained the model to operate directly with Thomson Reuters’ internal research tools.

The total build cost was approximately $40 million, covering both personnel and compute over two years. The final training run for the version shipping today cost $450,000 — a figure Thomson Reuters is explicitly using to contrast with the billions frontier labs spend on comparable general-purpose models.

The foundational research team traces back to Safe Sign, an AI safety company Thomson Reuters acquired in August 2024 in what the company described as its first pre-revenue acquisition.

What it does

Thomson 1.0 launches in a single, narrow deployment: tabular analysis inside CoCounsel Legal. That feature handles structured document review — contracts, filings, regulatory submissions — where volume is high and domain specificity matters. Thomson Reuters’ CTO Joel Hron said the choice was deliberate: high-volume structured tasks are where a purpose-built domain model shows immediate cost and quality advantages over a general frontier model.

The new generation of CoCounsel Legal itself launched last Thursday. The platform, which was in beta since April and early access since June, is built on Anthropic’s Claude Agent SDK and designed for multi-step agentic workflows combining research, drafting, and verification.

Thomson 1.0 does not replace Anthropic’s Claude in CoCounsel — it takes over one specific task. Thomson Reuters says it intends to expand Thomson’s scope over time as the model improves, and that the strategy is deliberate: let its own model compound in quality through real-world usage data rather than licensing general capabilities at frontier cost.

Why now

Thomson Reuters currently relies on Anthropic and other frontier labs for much of CoCounsel’s intelligence layer. According to Business Insider, Thomson 1.0 was built explicitly to reduce that dependence — and cost. Hron confirmed the model is intended to handle areas where Thomson Reuters can outperform foundation models on domain-specific tasks while paying less per inference.

The proprietary content moat is the bet. Thomson Reuters controls Westlaw, one of the most comprehensive legal databases in existence. Hron’s argument is that moving a model’s frontier in a specialized domain requires expert data that other labs cannot replicate — not more general compute.

Thomson Reuters says Thomson has so far been trained on less than 10% of its total proprietary content.

What comes next

Thomson 1.0 is the first in a planned family of models carrying the Thomson name. A small variant is being released on Hugging Face under a non-commercial academic license for external evaluation. A full technical report, including standard public AI benchmarks and legal-specific evaluations, is expected this week. Thomson Reuters has also given a group of legal and AI academics direct access to the model for independent evaluation.

A developer API portal is in early development, with Hron demonstrating a draft version during a media briefing but cautioning it is not yet production-ready.

Key numbers

• Training investment: $40M (talent + compute over 2 years)
• Final training run cost: $450,000
• Foundation: Qwen 3.5 (open-weight)
• Training data: Westlaw, Practical Law, Checkpoint, Reuters
• First production use: tabular analysis in CoCounsel Legal
• Open model: small variant on Hugging Face (non-commercial academic license)